This study was conducted in the Department of Affective Disorders, Institute of Psychiatry and Neurology in Warsaw and in a center specializing in clinical trials between September 2017 and December 2018. Patients were enrolled from both inpatient and outpatient settings to capture both mild and severe episodes. Based on previous studies [28,29], it was expected that patients would change the phase at least once in the period envisaged for this study. This study obtained the consent of the Bioethical Commission at the District Medical Chamber in Warsaw (agreement KB/1094/17).

We used the following inclusion criteria: adults aged 18 years or older, who gave their informed written consent to participate in this study, diagnosed with BD, with at least two changes of phase in the last 12 months, and using a smartphone with internet access daily or declaring their willingness to use it for the study period. Exclusion criteria consisted of serious hearing problems and speech disorders (eg, dysarthria, aphasia).

The resulting sample included 84 eligible patients diagnosed with BD (according to the International Classification of Diseases, tenth revision [ICD-10]). The baseline sociodemographic and clinical characteristics of the final study sample are presented in Table 1.

The assessment of the mental state was carried out by psychiatrists with experience in the diagnosis and treatment of BD. Both the researcher and the patient were blinded to the data automatically collected by the BDmon app. Patients were invited to visit the researcher at least every 3 months. In the case of a suspected change in the mood state, patients were invited for an additional intervention visit. During personal visits, the primary outcome measures (17-point version of HDRS and YMRS) were used. In this paper, the assessment of HDRS and YMRS is used as binding to assess the severity of symptoms and the corresponding BD states.

The data concerning clinical and demographic characteristics were collected as a part of the initial visit. During the following visits, the patients were inquired about significant changes in the general health status and other important life events that might affect their activity and behavior. The patient’s mood state was also assessed using fortnightly phone-based interactions with the researcher. The basic and the most important objective of the phone-based assessment was to determine whether any significant change in the mood state was likely to occur since the previous contact with the patient. If so, the patient was invited for an intervention visit, and a full assessment using HDRS and YMRS was conducted. To shorten the time of the fortnightly calls, the researcher read to the patient questions prepared in the telephone questionnaire form (see Multimedia Appendix 1).

Based on HDRS and YMRS, the mood state was classified into 1 of the 4 phases (depression, hypomania/mania, euthymia, or mixed state). In previous studies, researchers have adopted different cutoff points on the HDRS and YMRS for training classifiers [11,12,30]. High thresholds of 13 points on the HDRS and YMRS [12] increase the validity of the diagnosis, focusing on more severe symptoms. The consequence of this approach might be counting the milder symptoms of a given phase to euthymia. We aimed at the identification of early symptoms of phase change and hence adopted lower cutoff points. The state of euthymia was defined as HDRS<8 and YMRS<6, depression as HDRS≥8 and YMRS<6, hypomania/mania as HDRS<8 and YMRS≥6, and mixed state as HDRS≥8 and YMRS≥6.

Smartphones collect behavioral data continuously whereas clinical assessments are much less frequent. Supervised learning requires both data types to be present. Consequently, we extrapolated the psychiatric assessment to 7 days before and 2 days after a visit. In the final sample, we included only patients with automatically collected smartphone data within this period. The asymmetry of the time window was inspired by Muaremi et al [31] who argue that a visit is typically related to some medical intervention, which might lead to a change in the patient’s mental state. The time frame before the visit differs across studies, for example, Faurholt-Jepsen et al [11,12] adopted a time span of the clinical assessment days and 3 previous days. In this study, we extended this period up to 7 days before the visit because we had additional data about patient’s mental state (derived from phone-based visit and from patients’ relatives). Nevertheless, to facilitate meta-analyses, we additionally performed the analysis for the shorter period assuming ground truth of 3 days before the psychiatric assessment and the day of psychiatric assessment, as well for different cutoff points (Multimedia Appendix 2 and Multimedia Appendix 3).

The BDmon app requires a smartphone with the Android system, which at the time of the clinical trial had over 90% of the market share in Poland. Patients used their own smartphones or were offered a new smartphone for the study period. Patients did not receive any financial gratification for participation in the study. The BDmon app collected 3 groups of data:

(1) objective data about patient’s behavior related to smartphone (statistics about phone calls and text messages), (2) self-reported data completed by patients via the app (filling in the questionnaires was not mandatory; Multimedia Appendix 4), and (3) objective acoustic features about patient’s speech extracted from his/her daily phone calls. Only self-assessment (2) required the patients to interact with the software; data from groups (1) and (3) were collected passively. The scope of monitoring was as follows:

To assess the strength of the relation between the behavioral markers and the affective symptoms and BD states, we applied generalized linear mixed-effects models similarly as in the work of Faurholt-Jepsen et al [12]. First, mixed-effects linear regression was applied with scores on either HDRS or YMRS as response variables. This quantifies the correlations between behavioral markers and the severity of affective symptoms. Second, mixed-effects logistic regression was used for binary classification discriminating between an affective state and euthymia to describe the relation between the behavioral markers and affective states. The generalized linear mixed-effects model had the following form:

where y is an n_d × 1 vector consisting of n_d responses, each of which corresponds to 1 patient and 1 day. When assessing the severity of symptoms, either the HDRS or YMRS scores are used as the response vectors.

X is an n_d × s covariate matrix of the s predictor variables for fixed effects β (regression coefficient). The fixed effects consist of coefficients related to the objective smartphone data (eg, number of incoming calls) and a constant term. We build a separate model for each of the predictor variables; therefore, s=1.

Z is the n_d × q covariate matrix for the random effects u. The random effects occur at the patient level (level two) and are patient-specific random intercepts. Consequently, q is equal to the number of patients (groups) for whom data were available.

ε is the vector of n_d errors, which is assumed to be multivariate normal with zero mean. The number of patients differs depending on the considered parameter. For example, if patient A was never sending text messages, the daily number of sent text messages was calculated as 0. This allowed us to include patient A for the assessment of the relation between symptoms and the number of sent text messages. At the same time, sending no text messages led to missing values reported for the mean number of characters of sent text messages, and thus, patient A was not counted for assessment of the relation between symptoms and the length of sent text messages. Logistic mixed-effects regression models were applied to assess the relations between pairs of euthymia and the affective states (eg, euthymia vs depression). For each of the logistic models, the response variable equaled 0 for days with the euthymic state and 1 for days with the affective state (eg, depression). The inverse of the log link binomial function h(η) = (1+exp(η))⁻¹ was applied to the linear predictor η=Xβ + Zu to relate it to the outcome y. The mixed-effects logistic regression model was formulated as follows:

Model assumptions were checked with residual analysis and visually using quantile-quantile plots. The P values were calculated assuming normal distributions of errors. However, for some variables, especially the number of outgoing and incoming calls per day, this assumption seemed to be violated. Therefore, caution is necessary when interpreting the P values of the mixed-effects models; see also Luke [34]. Log transformations were applied to minimize this effect and the results were gathered for comparative purposes in Multimedia Appendix 5 and Multimedia Appendix 6. All analyses were conducted in the R programming language (R Core Team and the R Foundation for Statistical Computing). Linear and logistic mixed-effects models were calculated using packages lme (for the restricted maximum likelihood estimation) and lmerTest (P values and model diagnostics) available at the Comprehensive R Archive Network repository for R language [35]. We set the significance level to .05.

Figure 1 illustrates the participant flow in this study; 84 eligible patients diagnosed with BD (according to ICD-10 classification) were enrolled in this study, and they participated in the initial interview with the psychiatrist. Disregarding their declarations of participation in this study, 20 patients dropped out after the initial psychiatric assessment. As depicted in Figure 1, the mean age (36.6 [SD 9.1] years) was lower for the group of 64 patients who continued the study in comparison to that of the group of 20 patients who resigned after the initial visit (44.3 [SD 14.6] years). For the remaining 64 patients, there were 226 psychiatric assessments in total. However, smartphone-based automatically collected behavioral data in the assumed time frame (7 days before and 2 days after psychiatric assessment) were available for only 51 patients (196 psychiatric assessments). These patients (N=51) were considered as the final study sample. In total, this constitutes 982 patient days with data for the statistical analyses. The mean participation time for the final sample of 51 patients, calculated as the difference between the initial and the last psychiatric assessment, was 208 days with a standard deviation of 132 days.

Overall, 196 psychiatric assessments were reported for the final study sample. We observed 145 mood state transitions among consecutive psychiatric interviews as summarized in Table 2.

Out of 145 transitions, in 74 (51%) cases, the affective state changed from the previous one and in 71 (48.9%) remained the same. The most frequent change was from euthymia to depression state (17 such cases) and from depression to euthymia (14 cases).

Figure 2 illustrates the completeness of the objective behavioral data collected from smartphones. For each patient, there are 2 rows of dots. Green dots in the upper row show days, in which the objective behavioral data were collected. Turquoise dots in the lower row indicate days with self-assessment data. Less frequent usage of the app in early 2018 was a result of technical stability issues with the app.

Overall, we identified 982 person days with smartphone data and psychiatric assessments. The summary statistics of behavioral and self-assessment data are presented in Table 3 for all variables as initially planned for this study.

The app was designed to collect data concerning mobility (daily travelled distance) and the activity measures with the number of steps (pedometer), but owing to the technical issues and patients’ privacy concerns, the completeness of these data (98/982, 9.9%) was insufficient to conduct a reliable analysis. Similarly, the collected self-reported comments about sleep and mood were characterized by relatively high rates of missing data (n=39 and n=37, respectively). Table 4 provides the sociodemographic and clinical characteristics of patients depending on their adherence in terms of filling in self-assessment questionnaires. The study patients were split into 3 groups depending on the completeness of the self-assessment data; 16 patients out of 51 patients (31%) demonstrated low adherence to filling in self-assessment questionnaires (completeness less than 98 patient days). The group of patients with low adherence had relatively high sum of points on the HDRS (7.25 [SD 5.05]) and low sum of points on the YMRS (1.81 [SD 1.64]). Interestingly, 10 patients out of 16 in the group with low adherence (62.5%) were assessed as euthymic during the initial assessment.

The summary statistics about the behavioral and self-assessment data in various affective states are shown in Multimedia Appendix 7. For most of the considered variables, differences in statistics were observed between the 4 groups (euthymia, depression, mania, mixed state). Apart from the overall difference of statistics for patients in various BD states, there was also a relatively high intrapatient variability in phone call statistics. Figure 3 illustrates this for 2 objective parameters and all patients with BD. The boxplots for all the variables are presented in Multimedia Appendices 8-19.

Table 5 presents the estimates of the regression coefficients β. They describe the influence of behavioral data on the severity of manic and depressive symptoms measured by the number of HDRS and YMRS points. The number of observations is the product of the number of groups (ie, patients) and days for which the considered data type was collected.

The results show that the more severe the manic symptoms: (1) the higher variability in the duration of incoming calls (P=.04) and outgoing calls (P=.02), (2) the higher number and fraction of missed calls (P=.02), (3) the higher number of outgoing text messages (P<.001), (4) the higher number of characters in outgoing text messages (P<.001), and (5) the higher score on self-reported mood rating (P<.001). For example, for every increase of 1 missed call, there was an increase of 0.141 points on the YMRS. More severe depressive symptoms are associated with (1) lower number of incoming answered calls (P<.001), (2) higher fraction of incoming missed calls (P<.001), and (3) lower score on self-reported mood rating (P<.001).

Tables 6 and 7 present the results from generalized mixed regression models (logistic model). Smartphone-based data were used as predictors. We were interested in the detection of behavioral changes between euthymia, which we treated as a reference and encode as 0, and the other affective states encoded as 1. Therefore, for each predictor, we fitted 4 regression models distinguishing euthymia from depression, mania, mixed state as well as a combination of manic and mixed states. Positive values of the regression coefficients indicate that the predictor tends to have lower value in the reference euthymic state. As we put emphasis on the early identification of phase change by adapting lower cutoff points on the above scales, we were able to distinguish between euthymia and other affective states, both mild and severe, that is, depression, hypomania/mania, or mixed state.

The following variables discriminate between euthymia and depression: (1) number of incoming answered calls (P=.01), (2) fraction of missed calls (P<.001), and (3) number of outgoing calls (P=.01). Euthymia and mania differ significantly in fraction of outgoing calls (P=.03). The mixed states observed in this study were mainly mixed manic states, which was reflected in the average scores on HDRS (10.6 [SD 2.9]) and YMRS (17.6 [SD 10.2]). This is the rationale behind conducting additional analysis to catch the whole spectrum of manic or mixed features by combining both states into 1 group in Table 7. The following variables turned out to be relevant and discriminate the above conditions from euthymia: (1) fraction of missed calls (P=.01), (2) variability of the duration of calls (P=.045), (3) number of sent text messages (P=.01), and (4) mean length of text messages (P=.01).

Further, in Table 8, we discriminate euthymia from the pathological states considered as 1 group to find out which predictors are relevant markers for BD. The following 7 out of 13 variables (54%) are statistically significant: (1) number of incoming calls (P=.002), (2) fraction of outgoing calls (P=.01), (3) fraction of missed calls (P<.001), (4) number of sent text messages (P=.01), (5) mean length of text messages (P=.01), (6) self-reported sleep time (P=.03), and (7) self-reported mood (P=.003). It needs to be noted that 6 out of these 7 variables (all except for the self-reported sleep time) are statistically significant and discriminate euthymia from the pathological states considered as 1 group also for the longer period (14 days) preceding the psychiatric assessments (see Multimedia Appendices 8-20).

The most clinically relevant data that could be drawn from both presented analysis types are as follows: (1) number of incoming answered calls was lower in patients with depression as compared to those with euthymia and, at the same time, missed incoming calls were more frequent and increased as depressive symptoms intensified; (2) depressed patients tended to make phone calls less frequently than euthymic patients; (3) fraction of missed calls was higher in manic/mixed states and was positively correlated to manic symptoms; (4) fraction of outgoing calls was higher in manic states; (5) variability of duration of calls was higher in manic/mixed states and positively correlated to the severity of symptoms; and (6) the number and length of sent text messages was higher in manic/mixed states as compared to euthymic state and positively correlated to the severity of manic symptoms. The self-reported mood scores were significantly correlated with depressive and manic symptoms (as measured with the HDRS and YMRS), but data were insufficient to explain their relation to BD states.

To the best of our knowledge, this is one of the largest studies investigating the relation between the BD phase assessed by psychiatrists and the objective behavioral data collected via smartphones. Reliable smartphone-based results were obtained for 51 patients. This paper confirms and further extends the findings presented by Faurholt-Jepsen et al [12] concerning objective behavioral markers based on phone call statistics. Our results confirm that the majority of phone call statistics are significantly correlated with the clinically rated depressive and manic symptoms (assessed using the HDRS and YMRS, respectively) and are valid markers distinguishing between BD phases (mania, depression, mixed state, euthymia). We observed that depressed patients made phone calls less frequently than euthymic patients; they also answered the phone less often, and this behavior increased as the depressive symptoms intensified. However, patients in the manic/mixed states used phones more frequently than patients in the euthymic state, which was seen in the higher fraction of outgoing calls and a higher number and length of text messages. In general, these results are consistent with those reported in a previous research. For example, in the studies of Faurholt-Jepsen et al [8] and Muaremi et al [31], the number and length of phone calls were correlated with the BD phase and were found to be higher in mania and lower in depression. However, the available studies are not entirely homogeneous in terms of the given phone call parameter and its direction of changes in the particular phase of BD [8,12,31,36]. Patients present different behavioral patterns depending on individual characteristics and preferences for using a mobile phone. We suppose that this issue may at least partly depend on the analyzed group of patients and intrapatient variability. In our study, we also observed that patients in manic/mixed states demonstrated various phone usage patterns—from very long calls to very rare phone use. This was especially reflected in the large standard deviation of the duration of calls. To our knowledge, this variability has not been investigated in detail in similar studies so far. This could open up an interesting field for further research focused on tailored solutions for the individual patient rather than on the characteristics of BD itself. It seems that taking precautions in considering call statistics as a single-phase marker or phase change marker for all patients, in particular, in manic/mixed state, would be reasonable. The behavioral parameters differ between mood phases (depression, mania, mixed state) as compared to euthymia. Depression and mania reflect various manifestations of the illness, but not necessarily should be seen as two poles of the same dimension [37]. Consequently, parameters characterizing depression and mania do not need to have opposite signs but will rather be different from euthymic state. In our study, this conclusion was also confirmed by the observation of missed calls among patients with depression and mania or mixed state. This parameter was significantly increased in both affective states as compared to euthymia and was positively correlated to the severity of depressive and manic symptoms. This could indicate that both manifestations of the disease, although of different affective tones, have a similar effect on social interactions, and its size increases with the severity of the illness symptoms. In depression, this could be due to psychomotor retardation and social isolation, while in manic/mixed states, this could be the effect of growing distractibility and inattention. Therefore, we conjecture that the key information might be conveyed by the general change in behavioral patterns appropriate for a given patient in euthymia, and not the direction of the change itself. The above conclusion is also supported by an analysis comparing all pathological affective states in BD to the euthymic state (Table 8), showing a significant difference with the euthymic state. The correlation of self-assessment data with the clinically rated depressive or manic symptoms has been shown in the previous research [8,38-41]. It is worth noting at this point that there are also apps providing personalized psychoeducation programs or mobile therapy combined with self-assessment tools, for instance, the SIMPLe app [42]. The SIMPLe app showed efficacy in improving sleep, social rhythms, and eating pattern [43].This paper also revealed that in natural settings, patients’ adherence declined over time. We observed that at the very beginning, almost 90% of the patients filled data regularly, but after 3 months, only less than one-quarter of the patients systematically continued the self-assessment. This adherence rate appeared to be similar to that reported by Hidalgo-Mazzei et al [44] where only 30% of the patients were using the app regularly after 6 months of the study. Nevertheless, it should be stressed that adherence in completing self-rated questionnaires ranging from 42% to 95% was observed in previous studies [45]. We have identified that most of the patients with low adherence in filling self-assessment questionnaires were in the state of euthymia during the initial psychiatric assessment. The other hypothesis is that cognitive performance might play a certain role in filling the assessments [46]; yet, the data collected in this study provides little information on this topic. Further research might take this potential variable into consideration. The long-term collection of self-assessment data and their suitability for the recognition and prediction of affective state seem to be problematic. Similar conclusions were also drawn by Faurholt-Jepsen et al [8]. The completeness of the self-assessment data collected in this study was insufficient to fully explain their relation to BD states. A strength of this study is that most patients were assessed several times using a longitudinal design for an average of ~7 months. The adopted cutoff points on the depressive (HDRS) and manic (YMRS) scales allowed us to capture changes in phone usage patterns that are already different in mild depression/hypomania/mania/mixed states as compared to euthymia. Patients with severe manic/mixed symptoms also tended to provide more diversified patterns of behavior in the same condition. This sheds new light on, so far little studied, the variability of behavior among patients with BD (measured for example with the standard deviation) in given affective states. Further research could concentrate on personalized apps, adapting to a given patient, or the search for more generalizable smartphone-based objective parameters, which will be independent of the individual patterns of the phone usage and behavior. Acoustic features of the human voice seem to be promising candidates [47]. In the study of Faurholt-Jepsen et al [11], voice data turned out to be a reliable objective phase marker in BD. Moreover, speech features seem to be a promising marker for the assessment of suicide risk in patients with depression [48]. In our study, we also collected these data, but owing to their extensive nature and the need for thorough analysis, they will be presented in a separate paper. Relations between data collected from smartphones and the depressive and manic affective symptoms confirmed in this study demonstrate promising potential for both early detection of affective states and the prediction of phase changes. This is commonly formulated as either classification or regression task [49]. Recent papers [34,50] show that statistical and machine learning approaches can be complemented by process monitoring with control charts. These are easily understandable visual tools that naturally address the temporal structure of data and generate notifications about the change of BD phase.

The initially planned sample size was 100 patients. The reason for not meeting this goal was a slowdown in the trial in early 2018 due to technical issues (eg, monitoring of the daily travelled distance based on satellite navigation data, disruption of the normal usage of smartphone). However, it is worth noting that the final sample size (N=51) is still one of the largest among similar studies. The BDmon app was developed only for Android as it was the dominating operating system in Poland at the time of this study and we did not want to add additional costs and technical complexity to the project. During the enrolment into the study, only 3 out of 84 patients (4%) did not have an Android smartphone and were offered a relevant device for the study period. Therefore, we believe that the potential autoselection of our sample arising from this issue was negligible. The most controversial monitoring function for patients was travelled distance computed using satellite navigation—almost 90% of the patients refused to be monitored; thus, these data were not eventually analyzed. Moreover, patients in the manic state were likely to switch off the smartphone quite frequently or uninstall the app, which created nonrandom missing data. During this study, 3 patients experienced worsening of the psychotic symptoms accompanying depressive and manic episodes. Finally, since this study was exploratory, no correction for multiple comparisons was applied.

This study brings strong evidence that smartphone-based parameters reflecting behavioral activities are related to the severity of depressive and manic symptoms and allow for discriminating between affective states in BD, that is, depression versus euthymia and manic/mixed states versus euthymia. These parameters could be used to assess the severity of manic or depressive symptoms in BD and assist in the early recognition of phase change, which can increase the patient’s chance of early intervention between outpatient visits and hence improve the course of the disease and prognosis.